CanID: a robust and accurate RNAseq Expression-based diagnostic classification scheme for pediatric malignancies

Area of Science:

• Oncology
• Bioinformatics
• Machine Learning

Background:

• Pediatric cancer subtype classification is crucial for targeted therapies.
• Current methods face challenges in scope, precision, and standardization for pediatric cancers.
• Omics-based machine learning classifiers are increasingly used to augment histopathology.

Purpose of the Study:

• To develop a robust machine learning classification scheme for pediatric cancer subtypes.
• To address analytical challenges in classifying diverse pediatric solid tumors and hematologic malignancies.
• To provide a transcriptome-based approach for improved tumor diagnosis and stratification.

Main Methods:

• Developed Cancer Identification (CanID), a stacked ensemble machine learning model.
• Utilized gene-level RNA sequencing count data (transcriptomic features) as the sole input.
• Trained CanID on 3203 pediatric cancer samples across 13 solid tumor and 38 hematologic malignancy subtypes.

Main Results:

• Achieved 99% accuracy for solid tumors and 92-93% accuracy for hematologic malignancies on independent test datasets.
• Demonstrated robustness against biological and technical variations, including data collection differences and class imbalance.
• Successfully classified challenging subtypes that are difficult for clinical histology evaluation.

Conclusions:

• CanID offers a highly accurate and robust transcriptome-based classification scheme for pediatric cancers.
• The model's biological interpretability aids in advancing tumor diagnosis and stratification.
• CanID represents a valuable tool for precision oncology and clinical decision-making.